Ticket #175 (closed defect: fixed)

Opened 5 years ago

Last modified 4 years ago

a bug in end game

Reported by: caiq Owned by: gnugo
Priority: normal Milestone: 3.7.12
Component: source Version:
Severity: normal Keywords:
Cc: patch: yes

Description

In the game attached, gnugo current cvs version made a mistake in end game in move 52. You can repeat it by invoking gnugo with "gnugo --mode gtp --level 11 --score aftermath --capture-all-dead --chinese-rules" and then send gtp commands loadsgf test001.sgf 52 genmove w

gnugo play at G5 and lost the game.

Attachments

test001.sgf Download (1.3 KB) - added by caiq 5 years ago.
bug175_add_regression_test.patch Download (882 bytes) - added by josephpiche 4 years ago.
Adds regression test
seki_tests_7_12.1 Download (2.9 KB) - added by gunnar 4 years ago.
Seki tests including followup positions.
gunnar_7_12.12.diff Download (1.5 KB) - added by gunnar 4 years ago.
Fix for seki:1224.
seki_patterns_7_12.1.diff Download (783 bytes) - added by gunnar 4 years ago.
Patterns to solve seki:1201,1202.

Regression Results

Attachment Rev. PASS FAIL Nodes Status
bug175_add_regression_test.patch Download never tested
gunnar_7_12.12.diff Download 2412 1 -0.42% -0.84% -0.17% details
seki_patterns_7_12.1.diff Download 2414 2 +0.00014% 0% 0% details

Change History

Changed 5 years ago by caiq

comment:1 Changed 4 years ago by gunnar

  • Component changed from source to regressions
  • Milestone changed from 3.7.11 to 3.8

The issue is that this corner position








looks like it can be turned into a seki by playing at 2-2. If that's indeed the case it should go into seki.tst, cf. testcases seki:901-913.

To solve the problem, the currently best approach is to do like patterns S12 and S13 in patterns.db.

comment:2 Changed 4 years ago by josephpiche

  • Component changed from regressions to patterns
  • patch set

This patch adds a regression test for each color to check this position. Under the current CVS, both tests fail.

Changed 4 years ago by josephpiche

Adds regression test

Changed 4 years ago by gunnar

Seki tests including followup positions.

Changed 4 years ago by gunnar

Fix for seki:1224.

comment:3 Changed 4 years ago by gunnar

  • Component changed from patterns to source
  • Milestone changed from 3.8 to 3.7.12

comment:4 Changed 4 years ago by gunnar

The patch gunnar_7_12.12 improves the search for moves to make seki out of a tactically critical string next to an alive dragon. This solves seki:1224 without any other regression breakage. It also tightens up the heuristics for deciding when it's worth calling the semeai reading, primarily by requiring small escape values for the involved dragons. This gives a nice reduction of owl nodes. 

seki:1224       pass
1 PASS (0 PASS, 1 pass)
Total nodes: 1697268511 3317095 12389214 (-0.42% -0.84% -0.17%)

Changed 4 years ago by gunnar

Patterns to solve seki:1201,1202.

comment:5 Changed 4 years ago by gunnar

  • Status changed from new to closed
  • Resolution set to fixed
Note: See TracTickets for help on using tickets.